dflash A/B outcome: eugr build rejects all draft attention backends under fp8 KV (FLASHINFER/TRITON_ATTN/B12X/FLASH_ATTN) — DFlash2-for-GLM lives in the kindling overlay stack; the card stack is MTP-only

recipes/glm53-flash-fp4local-2x-dflash/recipe.toml

@@ -128,3 +128,10 @@ container = "/root/.triton"
 [[extra_mounts]]
 host = "~/.cache/b12x"
 container = "/root/.cache/b12x"
+
+# A/B OUTCOME (2026-10-17): the eugr build (dev21830+g8244d1c44) rejects every
+# draft attention backend under fp8 KV — FLASHINFER/TRITON_ATTN/B12X/FLASH_ATTN
+# ("FP8 KV cache requires FA3 on SM90 or FA4 on SM100"; B12X additionally
+# rejects the draft's block size). DFlash2 for GLM lives in the kindling
+# overlay stack; the eugr/card stack is MTP-only for this checkpoint. Keep this
+# recipe for when the build's DFlash path learns the fp8-KV/block matrix.